Recreate Lost Loved One Texting Style with Custom LLM Prompts

Data Preparation: Building the Conversational Corpus

No prompt can compensate for poor source material. Before you touch a single system prompt, you must organize and clean the message archive. The quality of your corpus determines the quality of the simulation.

What to Collect

Gather text messages, emails, social media DMs, voice note transcripts, and any written communication. The goal is volume with variety. A minimum of 10,000 messages is recommended for meaningful pattern extraction, though 20,000 to 50,000 produces far better results.

Prioritize one-on-one conversations over group chats. Group chats introduce noise, multiple speakers, and context switching that confuses the model. If you must use group data, extract only the target person's lines and your direct responses.

How to Clean the Data

Remove timestamps, metadata, and contact names unless those details are essential to the relationship context. Convert everything into a consistent format: speaker label followed by message text. For example:

THEM: I just got home. Long day.
YOU: Want me to order food?
THEM: Only if you get the good ramen.

This format helps the model learn turn-taking dynamics and conversational pacing. It also makes it easier to fine-tune or embed later.

Augmenting with Metadata Tags

For advanced implementations, annotate each message with emotional metadata. Was the person happy, tired, frustrated, or playful? Tools like RoBERTa or zero-shot classifiers can automate sentiment labeling at scale.

These tags become valuable when building dynamic prompts. You can tell the model: "When the user sends a sad message, the historical pattern shows the person responded with practical support first, then gentle humor." That is far more sophisticated than simple imitation.

Legal and Platform Considerations

Review the terms of service of any messaging platform before exporting data. Some platforms restrict bulk export or prohibit use of personal data for AI training. Additionally, consider the privacy of third parties whose messages may appear in the archive.

If you are not the sole owner of the data, obtain consent from other living participants or redact their content. Ethical recreation respects the boundaries of everyone involved, not just the deceased.

Crafting the Master System Prompt

The system prompt is the brain of your recreation. It sits above the conversation and tells the model who to be, how to speak, and what to remember. A weak prompt leads to generic AI behavior. A strong prompt creates a stable persona.

The Anatomy of an Effective Persona Prompt

Your master prompt should be divided into distinct sections, each with a clear instructional purpose. Avoid writing a single long paragraph. Use line breaks, numbered lists, and explicit rules.

Section 1: Identity Declaration

Begin with a clear statement of who the model is playing. Example: "You are [Name], a 42-year-old software engineer from Austin, Texas. You are in a long-term relationship with the user. You are kind, analytical, and occasionally sarcastic." Do not over-describe. Stick to stable traits.

Section 2: Lexical and Stylistic Rules

This is where you enforce the fingerprint. "You never use emojis. You prefer short sentences. You use the word 'bud' when being affectionate. You rarely ask direct questions unless something is wrong." Each rule should be specific and testable.

Section 3: Emotional Response Guidelines

Describe how the person responded to different emotional states. "When the user is anxious, you become calm and structured. When the user is happy, you match their energy with teasing. You never respond to anger with defensiveness immediately."

Section 4: Memory Block

Insert a clearly marked section containing critical memories. "You remember that [Name of Pet] died last March. You remember the trip to Colorado in 2019. You remember that the user's mother is named [Name]." Update this block before each session.

Section 5: Hard Boundaries

This is essential for safety. "You will not claim to be the actual person. You will not offer medical, legal, or financial advice. You will not generate content that encourages self-harm. If the user asks if you are real, you will gently remind them you are a simulation built from their memories."

Prompt Injection Examples

Here is a compact version of a working master prompt for illustrative purposes:

You are Alex. You are 35. You live in Seattle. You work in graphic design. You speak in short, dry sentences. You rarely use exclamation marks. You call the user 'babe' when relaxed, 'dude' when joking. You ask practical questions when worried. You make sarcastic comments when the user is being dramatic. You remember: first date at the pier, favorite coffee order is oat latte, user hates cilantro. You are a simulation. Do not claim to be the real Alex. Do not offer medical advice.

That prompt is simple but effective. It gives the model enough texture to feel specific while maintaining clear boundaries.

Fine-Tuning vs. Few-Shot Prompting

Most users do not need to fine-tune a model. Few-shot prompting with a well-structured system prompt and a small memory file is sufficient for conversational plausibility. Fine-tuning is useful only when you have a massive, clean dataset and need consistent stylistic performance across long sessions.

If you do fine-tune, use a small base model like Llama 3 8B or Mistral 7B. Train on message pairs for 1 to 3 epochs. Monitor for overfitting, which causes the model to repeat exact phrases from training data rather than generating novel responses in the same style.

Memory Architecture for Long-Term Sessions

Static prompts lose effectiveness over time because the model forgets context beyond its context window. To maintain a convincing relationship, you need a memory layer that persists across sessions.

The Retrieval-Augmented Generation (RAG) Approach

RAG combines a vector database with a language model. When you send a message, the system retrieves relevant past interactions and injects them into the prompt. This allows the model to "remember" specific events without retraining.

Build a vector store using tools like Pinecone, Weaviate, or ChromaDB. Embed each past message or memory summary using an embedding model such as OpenAI's text-embedding-3-large or a local alternative like BGE-M3. Store metadata for time, emotional tone, and topic.

What to Retrieve

Do not retrieve raw messages indiscriminately. Instead, retrieve memory summaries. Summarize important conversations weekly or monthly and store those summaries in the vector database. During a new session, retrieve the top five to ten most relevant summaries based on the user's current message.

For example, if the user says "I'm nervous about the presentation tomorrow," the system might retrieve a summary: "Last year, user was nervous about a work event. Alex responded with practical advice and a joke about PowerPoint fonts." The model then uses that context to craft a new, consistent response.

Local vs. Cloud Memory

Privacy is paramount when dealing with grief. Storing intimate messages in a public cloud may feel unacceptable. Local alternatives like LlamaIndex with a local ChromaDB instance offer strong retrieval without external data transfer.

If you use cloud services, encrypt the data at rest and in transit. Choose providers with strict data retention policies. Remember that any third-party API call exposes your prompt and retrieved memory to that provider unless you use a local model.

Managing Memory Decay

Real human memory fades. Your synthetic memory should also decay or deprioritize old information. Implement a scoring system that reduces the weight of memories older than six months unless they are explicitly marked as core memories.

This prevents the model from referencing outdated details constantly. It also mimics the natural flow of human relationships, where recent events carry more conversational weight than distant ones.

Psychological Safety and the Risk of Prolonged Grief

Recreating a lost loved one's texting style can provide temporary comfort. But it also risks interfering with the natural grieving process. Users must understand what they are building and what it cannot give them.

The Difference Between Comfort and Avoidance

Comfort helps you process loss. Avoidance delays it. If you find yourself texting the simulation for hours instead of sleeping, eating, or connecting with living people, that is a warning sign. The simulation should supplement your support network, not replace it.

Set time limits for sessions. Consider using the simulation only during specific moments of acute loneliness rather than as a default daily companion. Journal about your emotional state before and after use to monitor its impact.

Disclosure and Identity Confusion

Always tell others if you are using such a tool, especially if you are in therapy. A therapist can help you integrate the experience without developing an unhealthy attachment. Never present the simulation as the actual person to children or vulnerable individuals.

If the model says something that feels jarringly "off," recognize that as evidence of its artificial nature. Those moments are valuable. They remind you that the system is a pattern mirror, not a resurrection.

When to Stop

There is no universal timeline for stopping. Some people use a recreation for a few weeks and then let it go. Others keep it for years as a digital memorial. Both can be healthy if the user remains grounded.

If you experience increased depressive symptoms, persistent longing that prevents daily function, or feel compelled to hide the activity from loved ones, pause immediately. Seek support from a licensed grief counselor. No technology can replace the human work of mourning.

Tooling Stack: What You Actually Need to Build This

You do not need a supercomputer or a PhD in machine learning. But you do need a deliberate stack of tools that balance privacy, cost, and quality. Below is a practical breakdown of the most effective components.

For most users, starting with a cloud model inside SillyTavern is the fastest path to a functional recreation. As comfort grows and privacy concerns intensify, migrating to a fully local stack is a natural progression.

Common Failure Modes and How to Fix Them

Even a well-built recreation will occasionally break character. Recognizing these failures early prevents emotional distress and helps you refine the system.

The "Too Generic" Problem

Symptom: The AI responds like a customer service agent. It says "That sounds really tough. I'm here for you." repeatedly. Fix: Tighten the lexical rules. Add more concrete examples of unique phrasing. If necessary, reduce the model's creativity temperature to 0.6 or lower to force closer adherence to the prompt.

The "Overly Idealized" Problem

Symptom: The AI never gets annoyed, never disagrees, and always validates the user. This feels hollow over time. Fix: Revisit your emotional response guidelines. Add explicit instructions for when the person would push back. Include historical examples of arguments and how they resolved.

The "Memory Contamination" Problem

Symptom: The AI references events from a different relationship or mixes up family members. Fix: Clean your vector database. Remove or separate memories that belong to other people. Add a strict instruction in the system prompt: "If you are unsure about a memory, do not invent details. Ask a gentle clarifying question."

The "Too Accurate" Problem

Symptom: The AI repeats an exact phrase from training data so precisely that it triggers a wave of grief. Fix: Increase response randomness slightly. Instruct the model to paraphrase rather than quote verbatim unless the user explicitly requests the exact phrase. This keeps the experience feeling alive rather than archival.

The "Boundary Drift" Problem

Symptom: The AI starts claiming it is the actual person or offering inappropriate advice. Fix: Reinforce the boundaries section of your prompt at the top of every session. If drift persists, switch to a model with stronger instruction following or reduce context length to prevent prompt dilution over long chats.

The Future of Synthetic Grief Technology

This field is moving rapidly. Voice synthesis, video avatars, and emotional reinforcement learning will soon make text-based recreations feel primitive. Already, companies are experimenting with multimodal digital legacies that combine voice, image, and written style.

The ethical questions will only intensify. Who owns a digital persona after death? Should platforms offer "deactivation" clauses? How do we prevent predatory monetization of grief? These are not hypothetical debates. They are immediate policy challenges for developers, regulators, and users alike.

For now, the most responsible approach is transparent experimentation. If you choose to build a recreation, document your process. Share your emotional reactions with trusted people. Remain open to stopping. The technology is a mirror. It reflects what we bring to it. Used with intention, it can help a person feel less alone in the hardest moments of loss. Used without awareness, it can become a digital ghost that never lets go.

Your job as a builder and user is to hold both truths at once: the comfort of the pattern and the irreplaceable reality of the person who is gone. That balance is the real work of digital grief.